From 7b73f45541400db3deaa57c55a46e2c1d6398464 Mon Sep 17 00:00:00 2001
From: Samuel Oliveira Alves
 <107287165+SamuelOliveirads@users.noreply.github.com>
Date: Thu, 21 May 2026 02:11:17 -0300
Subject: [PATCH] Add adaptive sampling clone and free functions to manage
 memory (#1851)

---
 common/sampling.cpp  | 23 +++++++++++++++++++----
 src/llama-sampling.h |  4 ++++
 src/llama.cpp        | 12 ++++++++++++
 3 files changed, 35 insertions(+), 4 deletions(-)

diff --git a/common/sampling.cpp b/common/sampling.cpp
index ad5b7fae..99fc48ce 100644
--- a/common/sampling.cpp
+++ b/common/sampling.cpp
@@ -12,6 +12,9 @@
 #include <nlohmann/json.hpp>
 using json = nlohmann::ordered_json;
 
+struct llama_sampler_adaptive_p * llama_clone_adaptive_p(const struct llama_sampler_adaptive_p * adapt_p_ctx);
+void llama_free_adaptive_p(struct llama_sampler_adaptive_p * adapt_p_ctx);
+
 struct common_sampler * common_sampler_init(const struct llama_model * model, const struct common_params_sampling & params) {
     const llama_vocab * vocab = llama_model_get_vocab(model);
 
@@ -160,9 +163,11 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
             }
             case llama_sampler_type::ADAPTIVE_P:
             {
-                GGML_ASSERT(vocab);
-                auto n_vocab = llama_vocab_n_tokens(vocab);
-                result->adapt_p_ctx = llama_init_adaptive_p(n_vocab, params.adaptive_target, params.adaptive_decay, params.adaptive_updt_w_cur, result->rng());
+                if (params.adaptive_target >= 0.0f) {
+                    GGML_ASSERT(vocab);
+                    auto n_vocab = llama_vocab_n_tokens(vocab);
+                    result->adapt_p_ctx = llama_init_adaptive_p(n_vocab, params.adaptive_target, params.adaptive_decay, params.adaptive_updt_w_cur, result->rng());
+                }
                 break;
             }
             default:
@@ -185,6 +190,8 @@ void common_sampler_free(struct common_sampler * ctx) {
     }
     if (ctx->smpl)
         llama_sampler_dry_free(ctx->smpl);
+    if (ctx->adapt_p_ctx)
+        llama_free_adaptive_p(ctx->adapt_p_ctx);
     if (ctx->rbudget)
         common_reasoning_budget_free(ctx->rbudget);
     delete ctx;
@@ -255,6 +262,14 @@ void common_sampler_clone(common_sampler * src, common_sampler * dst) {
         dst->smpl = llama_sampler_dry_clone(src->smpl);
     }
 
+    if (dst->adapt_p_ctx) {
+        llama_free_adaptive_p(dst->adapt_p_ctx);
+        dst->adapt_p_ctx = nullptr;
+    }
+    if (src->adapt_p_ctx) {
+        dst->adapt_p_ctx = llama_clone_adaptive_p(src->adapt_p_ctx);
+    }
+
     if (dst->rbudget) {
         common_reasoning_budget_free(dst->rbudget);
         dst->rbudget = nullptr;
@@ -454,7 +469,7 @@ static void sampler_queue(
                     llama_sample_temp(ctx_main, &cur_p, temp);
                 }
                 break;
-            case llama_sampler_type::ADAPTIVE_P:  use_adaptive_p = true; break;
+            case llama_sampler_type::ADAPTIVE_P:  use_adaptive_p = ctx_sampling->adapt_p_ctx != nullptr; break;
             default : break;
         }
 
diff --git a/src/llama-sampling.h b/src/llama-sampling.h
index 80d5a4f2..3543df86 100644
--- a/src/llama-sampling.h
+++ b/src/llama-sampling.h
@@ -90,6 +90,10 @@ struct llama_sampler_adaptive_p * llama_init_adaptive_p_impl(int n_vocab,
         const bool updt_w_cur,
     const uint32_t seed);
 
+struct llama_sampler_adaptive_p * llama_clone_adaptive_p(const struct llama_sampler_adaptive_p * adapt_p_ctx);
+
+void llama_free_adaptive_p(struct llama_sampler_adaptive_p * adapt_p_ctx);
+
 void llama_prep_adaptive_p_impl(
               struct llama_sampling * smpl,
              llama_token_data_array * candidates,
diff --git a/src/llama.cpp b/src/llama.cpp
index 7651ec35..9d04f094 100644
--- a/src/llama.cpp
+++ b/src/llama.cpp
@@ -10533,6 +10533,18 @@ struct llama_sampler_adaptive_p * llama_init_adaptive_p(int n_vocab, const float
     return llama_init_adaptive_p_impl(n_vocab, target, decay, updt_w_cur, seed);
 }
 
+struct llama_sampler_adaptive_p * llama_clone_adaptive_p(const struct llama_sampler_adaptive_p * adapt_p_ctx) {
+    if (adapt_p_ctx == nullptr) {
+        return nullptr;
+    }
+
+    return new llama_sampler_adaptive_p(*adapt_p_ctx);
+}
+
+void llama_free_adaptive_p(struct llama_sampler_adaptive_p * adapt_p_ctx) {
+    delete adapt_p_ctx;
+}
+
 void llama_review_adaptive_p(struct llama_sampler_adaptive_p * adapt_p_ctx, const size_t n_unsent, const bool rewind_status) {
     llama_review_adaptive_p_impl(adapt_p_ctx, n_unsent, rewind_status);
 }